B a s i c S t a t i s t i c s F o r D o c t o r s Singapore Med J 2003 Vol 44(4) : 172-174
Randomised Controlled
Trials (RCTs) – Sample Size:
The Magic Number?
Y H Chan
INTRODUCTION
A common question posed to a biostatistician from
a medical researcher is “How many subjects do I
need to obtain a significant result for my study?”.
That magic number! In the manufacturing industry,
it is permitted to test thousands of components in
order to derive a conclusive result but in medical
research, the sample size has to be “just large enough”
to provide a reliable answer to the research question.
If the sample size is too small, it’s a waste of time
doing the study as no conclusive results are likely
to be obtained and if the sample size is too large,
extra subjects may be given a therapy which perhaps
could be proven to be non-efficacious with a smaller
sample size(1).
Another major reason, besides the scientific
justification for doing a study, why a researcher wants
an estimate of the sample size is to calculate the cost of
the study which will determine the feasibility of
conducting the study within budget. This magic
number will also help the researcher to estimate the
length of his/her study – for example, the calculated
sample size may be 50 (a manageable number) but
if the yearly accrual of subjects is 10 (assuming all
subjects give consent to be in the study), it will take
at least five years to complete the study! In that case
a multicentre study is encouraged.
STATISTICAL THEORY ON SAMPLE SIZE
CALCULATIONS
The Null Hypothesis is set up to be rejected. The
philosophical argument is: it is easier to prove a
statement is false than to prove it’s true. For example,
Clinical Trials and we want to prove that “all cats are black”, and even
Epidemiology if you point to me black cats everywhere, there’s
Research Unit
226 Outram Road still doubt that a white cat could be lying under a
Blk A #02-02
Singapore 169039 table somewhere. But once you bring me a white cat,
Y H Chan, PhD the hypothesis of ‘all cats are black’ is disqualified.
Head of Biostatistics Hence if we are interested to compare two
Correspondence to: therapies, the null hypothesis will be “there is no
Y H Chan
Tel: (65) 6317 2121 difference” versus the Alternative Hypothesis of
Fax: (65) 6317 2122 “there is a difference”. From the above philosophical
Email: chanyh@
cteru.gov.sg argument, not being able to reject the null hypothesis
does not mean that that it is true (just that we do not
have enough evidence to reject).
We want to reject the null hypothesis but could
be committing a Type I Error: rejecting the null
hypothesis when it’s true. In a research study, there’s
no such thing as “my results are correct” but rather
“how much error I am committing”. For example,
if in the population, there are actually no differences
between two therapies (but we do not know, that’s
why we are doing the study) and after conducting
the study, a significant difference was found which is
given by p<0.05.
There are only two reasons for this significant
difference (assuming that we have controlled for
bias of any kind). One is, there’s actually a difference
between the two therapies and the other is by chance.
The p-value gives us this “amount of chance”. If the
p-value is 0.03, then the significant difference due to
chance is 3%. If the p-value is very small, then this
difference happening by chance is “not possible” and
thus should be due to the difference in therapies
(still with a small possibility of being “wrong”).
The other situation is not being able to reject
the null hypothesis when it is actually false (Type II
Error). As mentioned, the main aim of a clinical
research is to reject the null hypothesis and we
could achieve this by controlling the type II error(2).
This is given by the Power of the study (1 – type II
error): the probability of rejecting the null hypothesis
when it is false. Conventionally, the power is set
at 80% or more, the higher the power, the bigger the
sample size required.
To be conservative, a two-sided test (more sample
size required) is usually carried out compared to a
one-sided test which has the assumption that the
test therapy will perform clinically better than the
standard or control therapy.
SAMPLE SIZE CALCULATIONS
To estimate a sample size which will ethically
answer the research question of an RCT with a reliable
conclusion, the following information should be
available.
 Singapore Med J 2003 Vol 44(4) : 173

Type of comparison(3) Effect size of therapies

Superiority trials
To show that a new experimental therapy is superior
to a control treatment
Null Hypothesis: The test therapy is not better
than the control therapy by a clinically relevant amount.
Alternative Hypothesis: The test therapy is better
than the control therapy by a clinically relevant amount.
Equivalence trials
Here the aim is to show that the test and control
therapies are equally effective.
Null Hypothesis: The two therapies differ by a
clinically relevant amount.
Alternative Hypothesis: The two therapies do not
differ by a clinically relevant amount.
Non-inferiority trials
For non-inferiority, the aim is to show that the new therapy
is as effective but need not be superior compared to the
control therapy. This is when the test therapy could be
cheaper in cost or has fewer side effects, for example.
Null Hypothesis: The test therapy is inferior to the
control therapy by a clinically relevant amount.
Alternative Hypothesis: The test therapy is not inferior
to the control therapy by a clinically relevant amount.
A 1-sided test is performed in this case.
Type of configuration(4)
Parallel design
Most commonly used design. The subjects are
randomised to one or more arms of different therapies
treated concurrently.
Crossover design
For this design, subjects act as their own control,
will be randomised to a sequence of two or more
therapies with a washout period in between therapies.
Appropriate for chronic conditions which will return
to its original level once therapy is discontinued.
The effect size specifies the accepted clinical difference
between two therapies that a researcher wants to
observe in a study.
There are three usual ways to get the effect size:
a. from past literature.
b. if no past literature is available, one can do a
small pilot study to determine the estimated
effect sizes.
c. clinical expectations.
To calculate the sample size, besides knowing
the type of design to be used, one has to classify the
type of the primary outcome.
Proportion outcomes
The primary outcome of interest is dichotomous
(success/failure, yes/no, etc). For example, 25% of the
subjects on the standard therapy had a successful
outcome and it is of clinical relevance only if we
observe a 40% (effect size) absolute improvement
for those on the study therapy (i.e. 65% of the
subjects will have a successful outcome). How
many subjects do we need to observe a significance
difference?
For a two-sided test of 5%, a simple formula to
calculate the sample size is given by
π1(1 – π1) + π2(1 – π2)
m (size per group) = c X
(π1 – π2)2
where c = 7.9 for 80% power and 10.5 for 90%
power, π1 and π2 are the proportion estimates.
Thus from the above example, π 1 = 0.25 and
π2 = 0.65. For a 80% power, we have
m (size per group) = 7.9 X [0.25 (1 – 0.25) +
0.65 (1 – 0.65)]/(0.25-0.65)2
= 20.49
Hence 21 X 2 = 42 subjects will be needed.

Type I error and Power(5) Table I shows the required sample size per group
The type I error is usually set at two-sided 5% and for π1 & π2 in steps of 0.1for powers of 80% & 90% at
power is at 80% or 90%. two-sided 5%.

Table I
π 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 199 (266) 62 (82) 32 (42) 20 (26) 14 (17) 10 (12) 7 (9) 5 (6)
0.2 – 294 (392) 82 (109) 39 (52) 23 (30) 15 (19) 10 (13) 7 (9)
0.3 – 356 (477) 93 (125) 42 (56) 24 (31) 15 (19) 10 (12)
0.4 – 388 (519) 97 (130) 42 (56) 23 (30) 14 (17)
0.5 – 388 (519) 93 (125) 39 (52) 20 (26)
0.6 – 356 (477) 82 (109) 32 (42)
0.7 – 294 (392) 62 (82)
0.8 – 199 (266)
Numbers in ( ) are for 90% power
174 : 2003 Vol 44(4) Singapore Med J

Continuous outcomes and unless an error message is obtained, it is most

Two independent samples
The primary outcome of interest is the mean difference
in an outcome variable between two treatment groups.
For example, it is postulated that a good clinical
response difference between the active and placebo
groups is 0.2 units with an SD of 0.5 units, how
many subjects will be required to obtain a statistical
significance for this clinical difference?
A simple formula, for a two-sided test of 5%, is
2c
m (size per group) = +1
δ2
µ2 - µ1
where δ = is the standardised effect size and
σ mention a few. Thomas & Krebs(7) gave a review of
µ1 and µ2 are the means of the two treatment groups
σ is the common standard deviation
c = 7.9 for 80% power and 10.5 for 90% power
From the above example, δ = 0.2/0.5 = 0.4 and for
a 80% power, we have m (size per group) = (2 X 7.9)/
(0.4 X 0.4) + 1 = 99.75
Hence 100 X 2 = 200 subjects will be needed.
Table II shows the required sample size per group
for values of δ in steps of 0.1 for powers of 80% &
90% at 2-sided 5%
Paired samples
In this case, we have the pre and post mean difference
of the two treatment groups and a simple formula is
Total sample size = c 1. Fayers PM & D Machin. Sample size: how many subjects patients
+2
δ2
Table III shows the total size required for values
of δ in steps of 0.1 for powers of 80% and 90% at
two-sided 5%.
SAMPLE SIZE SOFTWARE
There are many sample size calculations software
available in the Internet and even on most computers.
The main point to note in using a software is to
understand the proper instructions of getting the
sample size. One could enter some data into a program,
likely the magic number being generated is accepted
by the user. For this number to be “correct”, the right
formula must be used for the right type of design and
primary outcome. It is important to note that nearly all
the programs would provide the sample size for one
group and not the total (except for paired designs).
A simple-to-use PC-based sample size software,
affordable in cost, is Machin’s et al(6) Sampsize version
2.1 but it could only be installed for Windows 98 and
below. Software with network capabilities are SPSS
(www.spss.com), STATA (www.stata.com) and Power
& Precision (www.PowerAnalysis.com), just to
the various statistical power analysis software,
comparing the pros and cons.
CONCLUSIONS
This article has thus far covered the basic discussions
for simple sample size calculations with two aims in
mind. Firstly, a researcher could calculate his/her own
sample size given the types of design and measures of
outcome mentioned above; secondly, it is to provide
some knowledge on what information will be needed
when coming to see a biostatistician for sample
size determination. If one is interested in doing
an equivalence/non-inferiority study or with survival
outcomes analysis, it is recommended that a biostatistician
should be consulted.
REFERENCES
are necessary!. British Journal of Cancer 1995; 72:1-9.
2. Muller KE & Benignus VA. Increasing scientific power with
statistical power. Neurotoxicology & Teratology, 1992; 14:211-9.
3. Schall R, Luus H & Erasmus T. Type of comparison, introduction to
clinical trials, editors Karlberg J & Tsang K, 1998; pp:258-66.
4. Chan YH. Study design considerations — study configurations,
introduction to clinical trials, editors Karlberg J & Tsang K. 1998;
pp:249-57.
5. Thomas L & Juanes F. The importance of statistical power analysis:
an example from animal behaviour. Animal Behaviour, 1996; 52:856-9.
6. D Machin, M Campbell, Fayers P & Pinol A. Sample size tables for
clinical studies, 2nd edition. Blackwell Science, 1997.
7. L Thomas & CJ Krebs. A review of statistical power analysis
software. Bulletin of the Ecological Society of America 1997, 78(2):
126-39.

Table II
δ
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
80% power 1,571 394 176 100 64 45 33 26 21
90% power 2,103 527 235 133 86 60 44 34 27

Table III
δ
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
80% power 792 200 90 52 34 24 19 15 12
90% power 1,052 265 119 68 44 32 24 19 15